Research On Dereverberation Algorithms Of Array Microphones In Noise Environment

When people make phone calls,have a video conference or control the smart home through speech in offices and conferences or other enclosed environments,the speaker's voice is inevitably disturbedbyreverberation.Because when the microphone exceeds a certain distance from the sound source,the microphone usually collects direct sound and the reflect sound waves from the surrounding objects or walls at the same time,resulting in greater reverberation,and the reverberation degree is variable,which will lead to different degrees of speech intelligibility and clarity reduction,and will also affect people's auditory perception and user experience of related products.Therefore,speech dereverberation technology has important application value and research significance.This paper begins with a detail introduction of the research significance and current research situation of speech dereverberation;the basic principles of reverberation,including the generation,mathematical models and characteristics of reverberation;and the commonly used speech preprocessing operations;and the evaluation criteria of speech quality,and the experimental data in this paper.Then,the widely used weighted linear prediction error method is introduced in detail,and the application of multi-channels dereverberation is experimented.The results show that the performance of this algorithm decreases as the reverberation time becomes longer and longer.Owing to the process of this algorithm is equivalent to a linear filter,it will not cause speech distortion in processing.Since the estimation of reverberation components is inaccurate when the reverberation is serious,it still will make speech sound not smooth.Deep learning can better achieve the purpose of dereverberation through the non-linear computing,but the current common problem is that when the test set and training set do not match the reverberation environment,the performance of the algorithm will be seriously degraded.Based on this problem,this paper improves the algorithm in the field of image enhancement by combining the characteristics of speech signal,and proposesa residualaware reverberation combined with multi-stream densely connected dereverberation neural network.Moreover,the reverberation speech data set with reverberation degree classification label information are constructed.Experimental results show that his algorithm effectively improves the deficiencies of the current dereverberation algorithm with deep learning,and has obvious advantages over traditional algorithms in several speech evaluation indicators.